What Counts as Evidence in New-Business Acceptance Research: From Words to Behavior, and Then Generative AI

Do people who say they want it actually buy it?

Acceptance research is research that checks whether customers will accept a new business or product before it is fully built or launched. The traditional form is the concept test and the price survey, in which researchers show a product concept and ask whether people would buy it and how much they would pay. In new-business development, the same activity is also called customer discovery, demand validation, or product-market fit validation.

However, this form has a known weakness: the answers people give and what they actually do can differ. Schmidt and Bijmolt (2020) conducted a meta-analysis of 77 studies (47 papers, 115 effect sizes) that compared willingness to pay measured with hypothetical questions against willingness to pay measured when people actually paid. They reported that the hypothetical measure was on average 21% higher. In the same analysis, indirect methods such as conjoint analysis exceeded real values more than methods that ask directly. This result runs against the common assumption in marketing research that indirect methods are more accurate.

What, then, do new-business teams treat as evidence that a business will be accepted? This note draws on 6 public programs, 5 surveys by research firms and consultancies, 16 first-hand materials by the practitioners who developed the methods, and 14 academic works, and it checks figures and quotations against the full texts. For works whose full texts could not be obtained and which are therefore described from their abstracts, “abstract” is added after the year.

What does practitioner best practice treat as strong evidence?

When the first-hand materials of the practitioners who developed the methods are placed side by side, they agree that what customers do is stronger evidence than what they say. Osterwalder (2017) wrote that interviews and surveys are fast to set up but that what customers say and what they do are two different things, and that a simulated sale in which customers do not know they are part of an experiment can come fairly close to real purchasing. Gilad (2017) created a table for scoring the evidence behind an idea. In this table, one’s own conviction scores 0 points, anecdotes such as a request reported by sales score 20, customer evidence such as product data across several months scores 500, statistically significant test results score 2,000, and post-launch usage scores 10,000. These scores are Gilad’s heuristics and were not derived from measurement. Nevertheless, the orders of magnitude show how large he considers the gap between words and behavior to be.

The recommended way of asking also moves away from opinions and toward past behavior. Cagan (2021) responded to the objection that customers cannot tell us what to build by stating that what is being tested is whether they will buy the product if it is built. In the jobs-to-be-done interview, Moesta and Spiek ask people who recently switched to trace back the time leading up to the switch, and they write that jobs are not found in surveys or focus groups. According to them, a switch happens only when the push of the current situation and the pull of the new solution are strong enough to overcome the anxiety of the unknown and the habit of the present. For B2B new businesses in Japan, Hashiba (2026) wrote that the purpose of a hypothesis-testing interview is not to gather agreement but to face reality.

The order in which assumptions are tested is also specified by the methods. In the assumptions mapping of Bland (2020), a team writes out the assumptions that must be true for an idea to work, divides them into four types (desirability, feasibility, viability, and adaptability), and tests the important assumptions with little evidence first. Cagan (2017) also places value risk, that is, whether customers will buy the product or users will choose to use it, first among the risks of usability, feasibility, and business viability. Therefore, although acceptance is only one of the four types, it is listed first in both Bland’s and Cagan’s lists.

The expected frequency of this activity has also changed. Blank (2009) argued that there are no facts inside the building and urged founders to go outside and meet customers. Torres (2021) wrote that good product discovery teams engage with customers at least weekly and run interviews and assumption tests week over week, which fixed a frequency for this contact. As a result, acceptance research has come to be described as a habit continued at a fixed frequency while the product is being built.

There is also a measure of acceptance after launch. Vohra (2018) of Superhuman described how the company used Sean Ellis’s question, “How would you feel if you could no longer use the product?”, to measure the share of users who answer “very disappointed,” and reported that Ellis had set 40% as the threshold after benchmarking nearly a hundred startups. This 40% is Ellis’s comparison as relayed by Vohra, and the method of the comparison has not been published. In addition, the question asks for a feeling after use, so it remains stated evidence. Because it asks only users who have already started using the product, it belongs to a different stage from acceptance research before commercialization.

How do public support programs require validation of acceptance?

Among the public programs examined here, the US and the Japanese and European programs differ in the unit of what they require. The US NSF I-Corps program requires each participating team to complete at least 100 interviews with potential customers during its seven-week training (NSF 2022–2026). The 2025 solicitation (NSF 25-549) moved the training to an online-only format and expects teams to have completed 100 interviews with potential customers and partners by the end of the curriculum. The Department of Energy’s Energy I-Corps program challenges teams to conduct about 75 interviews during the training (DOE 2018).

I-Corps also counts the discovery that no market exists as an outcome. The NSF report of 2023 states that teams answer the product-market fit question with “Yes,” “No,” or “Not Yet” in their final presentation, and that even a “No” is deemed a success because it saves years of fruitless effort on a startup built around a technology without real market potential. This treatment is consistent with the experimental results discussed in the next section.

In Japan, NEDO’s support for commercializing new energy technologies (the application guidelines for fiscal year 2026) rejects proposals that lack a description of market needs and competing businesses, regardless of the results of the other review items. The EU’s EIC Accelerator requires applicants to show that the innovation has reached at least TRL 6 and to provide concrete and realistic figures on market size, market share, and sales price (EIC 2025). Both programs require evidence of the market as part of the application documents, but neither specifies how many customers to meet or how to validate. In short, the two US programs require the act of validation, whereas the two Japanese and European programs require a written description of its results.

The effects of these programs have not yet been established in the public bodies’ own reports. NSF (2023) reported that, from the start of I-Corps through the end of fiscal year 2022, more than 1,380 teams launched startups that subsequently raised 3.166 billion US dollars, 10.5 times the federal investment of 300 million US dollars. This is a cumulative count of participants’ outcomes, and it includes no comparison with researchers who did not participate. The Department of Energy (2018) attempted such a comparison but itself noted that the numbers may be misleading because the sample of non-participants was very small.

Do hypothesis-driven methods improve outcomes?

Academic research has carried out, with randomized experiments, the comparison that the program reports lack. Camuffo et al. (2020, abstract) divided 116 Italian startups into two groups and gave both groups ten training sessions on obtaining feedback from the market. Only one group was additionally taught to build a theory predicting the performance of its idea and to test its hypotheses rigorously. The abstract states that entrepreneurs who behaved like scientists performed better, were more likely to pivot to a different idea, and were not more likely to drop out in the early stages.

In a larger replication, the center of the results moved toward termination. Camuffo et al. (2024) studied 759 firms in four randomized experiments and reported that the treated firms terminated more ideas and ran a few radical pivots rather than none or repeated ones. Their supplementary analysis showed that this pattern of pivoting correlates with higher performance. Bailey et al. (2026, not yet peer reviewed) randomly assigned 553 science- and technology-based startups in 12 US co-working spaces to a training program that condensed an MBA entrepreneurship curriculum into six sessions. The treated startups shut down more often and sooner, and among the survivors they raised more external funding and had higher employment and revenue. This training did not teach acceptance validation alone. The authors interpret the result as the training leading entrepreneurs to reassess the expected value of their startups and to exit earlier.

Other studies show results in the same direction, under specific conditions. Koning et al. (2022, abstract) used the timing of adoption of A/B testing tools and state that, among the relatively few startups that adopted them, performance improved by 30% to 100% after a year, and that these startups scaled promising ideas and failed faster when they received negative signals. These data are observational, not randomized. Leatherbee and Katila (2020, abstract) followed 152 I-Corps teams and showed that the steps of formulating hypotheses, probing, and converging on a business idea link up as expected. The same abstract reports that team members with an MBA resisted the method and that, in a post hoc analysis, deeper engagement with the method related to higher performance in the following 18 months. Agarwal et al. (2025, abstract) compared, among agribusiness entrepreneurs in Tanzania, training that collects evidence only with training that first builds a theory and then collects evidence, and state that the latter produced higher economic performance.

This last result overlaps with a critique of the practitioner methods. Felin et al. (2020) argued that lean startup’s emphasis on readily observable feedback and immediately validated learning prompts a search for value and validation only where they are easy to observe, and undersells the task of composing a novel theory and hypotheses. Even after 100 interviews, a team whose hypothesis about what to test is ordinary learns only whether an ordinary idea is accepted.

The result shared by the two largest experiments was therefore the earlier termination of unpromising ideas, rather than better identification of good ones. The abstract of the first experiment with 116 firms states that early dropout did not increase, so the termination effect became clear only when the scale was expanded. The fact that I-Corps counts a “No” as a success points in the same direction as this result.

What is happening in new businesses at large Japanese firms?

In new businesses at large Japanese firms, few ideas pass through validation compared with the number of ideas generated. In a survey by PwC Consulting (2025) of companies with sales of at least 1 billion yen (n = 1,030), 79% of companies had their most advanced project reach the value validation or proof-of-concept stage, and 21% had a project that reached investment recovery. The same survey found that 40.8% of companies named internal idea generation as the measure they focused on most, whereas 9.0% named proof-of-concept experiments. NTT DOCOMO’s new-business program was reported to have produced 1,161 proposals, 115 validations, and 41 commercialization projects (Diamond Online 2022). The article does not describe what this “validation” consisted of.

Another survey asked about the relationship between the amount of validation and the outcome. The Innovation Research Institute (2026) surveyed 888 people who had been involved in new businesses at companies with at least 1,000 employees. It reported that the rate of reaching profitability (including the prospect of investment recovery) was 18.3% for businesses whose team talked directly with target customers 11 or more times, compared with 8.5% for those with 10 or fewer conversations. The institute itself notes that this shows only an association with the number of conversations and does not mean that more conversations will produce profitability. Because the survey was conducted by a firm that sells new-business support, the figures are read as claims of the source.

What stands out in the same survey is how negative evidence stops inside the organization. Of those who sensed that a project would fail, 90.9% told senior managers or stakeholders, but 16.6% of them were not heard. Among those who conveyed their sense of failure as the voice of actual customers, 23.6% were not heard. Acceptance research helps termination decisions only when the people who decide actually receive the negative evidence that has been collected. In new businesses at large firms, therefore, the way evidence is received matters in addition to the way it is collected.

What has generative AI changed in acceptance research?

Since generative AI spread, both ends of acceptance research have begun to change. The first change is on the building side. Blank (2026a) wrote that because teams can now build working products quickly with AI, minimum viable products (MVPs) no longer serve as proof of technical competence or as reliable evidence of customer discovery, hypothesis testing, or product-market fit. Blank renamed the MVP the “Initial Untested Product,” and in his class teams run weekly synthetic interviews with AI personas and compare those answers with in-person interviews (Blank 2026b). The same post states that more than 100 interviews exposed differences between users and buyers and between stated preferences and actual behavior. Because working prototypes have become cheap, evidence of acceptance has to be sought in customer behavior rather than in the quality of what has been built.

The second change is on the asking side. Synthetic respondents, that is, AI that generates answers in place of people, have entered market research. In a Qualtrics (2024) survey of 3,198 market researchers in 14 countries, 69% said they had used synthetic responses in the previous year, and 71% agreed that the majority of market research would use synthetic responses within three years (Research Live 2024). In the company’s survey in the third quarter of 2025, 45% of researchers who had adopted synthetic data named it as their most reliable data source (Qualtrics 2025). Both surveys were conducted by a company that sells synthetic response features, and they count the opinions and usage of researchers rather than measuring the accuracy of synthetic responses.

Studies that measured accuracy report that means come close but distributions do not. Argyle et al. (2023) showed that when GPT-3 is given the backgrounds of real survey participants, it can reproduce the response distributions of human subgroups in political surveys fairly closely, and they called this property algorithmic fidelity. Bisbee et al. (2024) reported that the mean answers of ChatGPT were close to a US election survey, but the answers had less variation, regression coefficients often differed significantly from the real data, and the distribution changed with minor changes in prompt wording. Wang et al. (2025) showed, with four language models and human studies involving 3,200 participants, that synthetic participants misportray demographic groups and flatten the diversity within them.

For price acceptance, the error of synthetic answers exceeded the error of human statements. Oetzel and Maiberger (2026) compared the willingness to pay of synthetic respondents with incentive-compatible real willingness to pay for four products. In all four studies the distributions differed significantly. For the two more hedonic products, synthetic respondents overstated willingness to pay by 86% and 37%, and the resulting profit forecasts were overestimated by 647% and 182%, respectively. The authors placed these errors against the 21% hypothetical overstatement of human answers (Schmidt and Bijmolt 2020) and concluded that synthetic respondents may support early pricing exploration but should not replace real data in price optimization.

There are also positive results. Maier et al. (2025, not yet peer reviewed) of PyMC Labs had a language model answer in free text and converted the semantic similarity of that text into five-point ratings. For 57 surveys by one company (9,300 human responses), they reported that this method reached 90% of human test-retest reliability. The authors belong to a company that works with synthetic responses, and a peer-reviewed version could not be confirmed. Sarstedt et al. (2024) reviewed how such results differ widely across domains and recommended silicon samples mainly for upstream parts of the research process, such as qualitative pretests and pilot studies.

Practitioners reach a similar position. Rosala and Moran (2024) of the Nielsen Norman Group state that synthetic users are useful for desk research and generating hypotheses but not for final decision-making. Torres (2024) wrote that she was disappointed to see the rise of generative AI tools designed to replace discovery with real humans. Synthetic Users, a company that sells synthetic users, claims 85% to 92% agreement with human research, but it has not published how this was measured or on what sample.

Synthetic respondents generate what people are likely to say. Human words already deviate from behavior, and synthetic answers deviate further from those words. In the order of evidence strength, it is therefore reasonable to place synthetic answers below interviews and questionnaires. The academic evaluations read here and the evaluations of practitioners who do not sell synthetic responses agree on a division of use: synthetic answers serve to expand hypotheses, and the decision on acceptance is confirmed with human behavior.

What does this mean as a procedure?

The materials above can be rearranged into a procedure for acceptance research. The sources for each step are given in parentheses.

1. Build the hypothesis first: Write down, as a theory for each idea, whose problem the idea solves, what the problem is, and why this idea solves it (Felin et al. 2020; Agarwal et al. 2025, abstract).
2. List the assumptions and set the order: Write out the assumptions about desirability, feasibility, viability, and adaptability, and test the important ones with little evidence first (Bland 2020; Cagan 2017).
3. Ask about past behavior: Instead of asking “Would you buy this?”, ask how people recently struggled, what they used, and why they switched (Moesta and Spiek; Hashiba 2026). Show several options rather than collecting agreement with one (Hashiba 2026).
4. Set the number of interviews in advance: I-Corps requires 100 interviews in seven weeks, and Energy I-Corps about 75 (NSF 2022–2026; DOE 2018). The number exists to reveal differences between users and buyers and between words and behavior (Blank 2026a).
5. Raise the strength of evidence step by step: After interviews and questionnaires, move to behavioral evidence such as simulated sales and pre-orders (Osterwalder 2017; Gilad 2017). Read price answers from hypothetical questions as likely to be higher than real values (Schmidt and Bijmolt 2020).
6. Use synthetic answers only up to hypothesis building: Use synthetic answers to broaden questions and hypotheses, and check them against human answers before deciding (Blank 2026b; Rosala and Moran 2024; Sarstedt et al. 2024).
7. Decide whether to continue, change, or stop: Record a “No” as an outcome (NSF 2023). Limiting pivots to a few was associated with better performance (Camuffo et al. 2024).
8. Build a route for negative evidence to reach decision makers: In organizations where a sense of failure conveyed as the voice of customers is not heard, the seven steps above do not lead to termination (Innovation Research Institute 2026).

What this review can and cannot claim

This review can claim that practitioner methods converge on treating behavioral evidence as stronger, that the weakness of stated evidence has a measured basis, that randomized experiments show the effect of hypothesis-driven training mainly as earlier termination, and that synthetic answers come close in means but miss in distributions and prices.

However, it cannot claim two things. First, no experiment was found that directly compared the business success of acceptance research using behavioral evidence with that of research using stated evidence alone. The experiments compared the presence or absence of hypothesis-driven training. Second, no survey was found that counted which acceptance research methods Japanese companies use and how often. The PwC survey measures the stages that projects reached, not the use of particular methods.

Related notes

• Conducting User Interviews: What Academia and Industry Share and Where They Diverge: A comparison of interviewing practice in academia and industry, which forms part of acceptance research
• A Genealogy of Practitioner Methods for Reframing Problems: Who Made Them, Traced to Their Origins: Practitioner methods for reframing problems, including the three strands of Jobs-to-be-Done
• Gate Design for New Businesses: Stopping What Should Stop Without Killing What Could Grow: Gate design that uses the results of acceptance research to decide whether to keep investing or stop

Unverified items

No unverified statement remains in the main claims. The full texts of the following works could not be obtained (publishers and institutional repositories refused automated retrieval), so they were used only within the scope of their abstracts: Camuffo et al. (2020), Koning et al. (2022), Leatherbee and Katila (2020), and Agarwal et al. (2025). Bailey et al. (2026) and Maier et al. (2025) are preprints, and peer-reviewed versions could not be confirmed. The new-business surveys by McKinsey and BCG, the CB Insights survey of startup failure reasons, and the study of purchase intentions by Morwitz et al. (2007) could not be reached in full text and were not used.

References

All sources were accessed on 2026-11-11. The ledger and verification records are in source/review/new-business-acceptance-research/industry.md.

Public bodies

• European Innovation Council. 2025. “Tips for Applicants: EIC Accelerator.” https://eic.ec.europa.eu/tips-applicants-eic-accelerator_en
• National Science Foundation. 2022–2026. “NSF I-Corps Teams: Information for Applicants.” https://www.nsf.gov/funding/initiatives/i-corps/national-teams-applicants
• National Science Foundation. 2023. I-Corps Biennial Report 2023. https://nsf-gov-resources.nsf.gov/2023-06/TIP_I-CorpsReport_2023_Final%206.21.2023.pdf
• National Science Foundation. 2025. “NSF National Innovation Corps Teams Program (NSF National I-Corps Teams), NSF 25-549.” https://www.nsf.gov/funding/opportunities/nsf-national-innovation-corps-teams-nsf-national-i-corps-tm/nsf25-549/solicitation
• U.S. Department of Energy, Office of Energy Efficiency and Renewable Energy. 2018. Energy I-Corps Year Two Impacts Report. https://www.energy.gov/sites/default/files/2022-03/energy-i-corps-year-two-impacts-report-final_031418.pdf
• New Energy and Industrial Technology Development Organization (NEDO). 2026. “Shin enerugī tō no shīzu hakkutsu, jigyōka ni muketa gijutsu kenkyū kaihatsu jigyō kōbo yōryō” [Application guidelines for the R&D program for discovering and commercializing new energy technology seeds] (in Japanese). https://www.nedo.go.jp/content/800052317.pdf

Research firms and consultancies

• Innovation Research Institute. 2026. “Shinki jigyō ni kansuru jittai chōsa” [Survey on new businesses] (press release, in Japanese). https://www.oricon.co.jp/pressrelease/3023816/
• PwC Consulting. 2025. “Shinki jigyō kaihatsu no torikumi ni kansuru jittai chōsa 2025” [Survey on new-business development initiatives 2025] (in Japanese). https://www.pwc.com/jp/ja/knowledge/thoughtleadership/2025/assets/pdf/new-business-development-survey-2025.pdf
• Qualtrics. 2024. “AI to Drive Massive Changes to Market Research in 2025, Qualtrics Report Says.” https://www.qualtrics.com/articles/news/ai-to-drive-massive-changes-to-market-research-in-2025-qualtrics-report-says/
• Qualtrics. 2025. “Research Teams Not Using AI Are Four Times More Likely to Lose Organizational Influence.” https://www.qualtrics.com/news/research-teams-not-using-ai-are-four-times-more-likely-lose-organizational-influence/
• Research Live. 2024. “Researchers Predict Synthetic Responses Will Dominate in Coming Years.” https://www.research-live.com/article/news/researchers-predict-synthetic-responses-will-dominate-in-coming-years/id/5131923

Practitioners

• Bland, David J. 2020. “How Assumptions Mapping Can Focus Your Teams on Running Experiments That Matter.” Strategyzer. https://strategyzer.com/library/how-assumptions-mapping-can-focus-your-teams-on-running-experiments-that-matter
• Blank, Steve. 2009. “Get Out of My Building.” https://steveblank.com/2009/10/08/get-out-of-my-building/
• Blank, Steve. 2026a. “AI Killed the MVP – Long Live the IUP.” https://steveblank.com/2026/09/25/ai-killed-the-mvp-long-live-the-iup/
• Blank, Steve. 2026b. “Lean LaunchPad: The Next Generation.” https://steveblank.com/2026/09/30/lean-launchpad-the-next-generation/
• Cagan, Marty. 2017. “The Four Big Risks.” SVPG. https://www.svpg.com/four-big-risks/
• Cagan, Marty. 2021. “Discovery Excuses.” SVPG. https://www.svpg.com/discovery-excuses/
• Diamond Online. 2022. “Shinki jigyō o shanai kara ryōsan suru ni wa?” [How to mass-produce new businesses from within: Lessons from Recruit, SoftBank, and NTT DOCOMO] (series “Daikigyō hakku taizen,” in Japanese). https://diamond.jp/articles/-/290883
• Gilad, Itamar. 2017. “Evidence Scores: The Acid Test of Your Ideas.” https://itamargilad.com/evidence-scores-the-acid-test-of-your-ideas/
• Hashiba, Eiko. 2026. “BtoB shinki jigyō no tame no ‘kokyaku intabyū’ 3 fēzu” [Three phases of customer interviews for B2B new businesses] (excerpt from Ichiji jōhō o “tou”, in Japanese). THE GOLD ONLINE. https://gentosha-go.com/articles/-/79459
• Moesta, Bob, and Chris Spiek. n.d. “Learn.” Jobs to Be Done. https://jobstobedone.org/learn/
• Osterwalder, Alexander. 2017. “How Strong Is Your Innovation Evidence?” Strategyzer. https://www.strategyzer.com/library/how-strong-is-your-innovation-evidence
• Rosala, Maria, and Kate Moran. 2024. “Synthetic Users: If, When, and How to Use AI-Generated ‘Research’.” Nielsen Norman Group. https://www.nngroup.com/articles/synthetic-users/
• Synthetic Users. n.d. Company website. https://www.syntheticusers.com/
• Torres, Teresa. 2021. “Continuous Discovery.” Product Talk. https://www.producttalk.org/continuous-discovery/
• Torres, Teresa. 2024. “Don’t Use Generative AI to Replace Discovery with Real Humans.” Product Talk. https://www.producttalk.org/generative-ai-discovery/
• Vohra, Rahul. 2018. “How Superhuman Built an Engine to Find Product/Market Fit.” First Round Review. https://review.firstround.com/how-superhuman-built-an-engine-to-find-product-market-fit/

Academic

• Agarwal, Rajshree, Francesca Bacco, Arnaldo Camuffo, Andrea Coali, Alfonso Gambardella, Haji Msangi, Steven Sonka, Anna Temu, Betty Waized, and Audra Wormald. 2025. “Does a Theory-of-Value Add Value? Evidence from a Randomized Control Trial with Tanzanian Entrepreneurs.” Organization Science 36 (2): 601–625. https://doi.org/10.1287/orsc.2023.17590
• Argyle, Lisa P., Ethan C. Busby, Nancy Fulda, Joshua R. Gubler, Christopher Rytting, and David Wingate. 2023. “Out of One, Many: Using Language Models to Simulate Human Samples.” Political Analysis 31 (3): 337–351. https://doi.org/10.1017/pan.2023.2
• Bailey, Esther E., Daniel C. Fehder, Eric J. Floyd, Yael V. Hochberg, and Daniel J. Lee. 2026. “Learning to Quit? A Multi-Year, Multi-Site Field Experiment with Innovation-Driven Entrepreneurs.” NBER Working Paper 34755. https://doi.org/10.3386/w34755
• Bisbee, James, Joshua D. Clinton, Cassy Dorff, Brenton Kenkel, and Jennifer M. Larson. 2024. “Synthetic Replacements for Human Survey Data? The Perils of Large Language Models.” Political Analysis 32 (4): 401–416. https://doi.org/10.1017/pan.2024.5
• Camuffo, Arnaldo, Alessandro Cordova, Alfonso Gambardella, and Chiara Spina. 2020. “A Scientific Approach to Entrepreneurial Decision Making: Evidence from a Randomized Control Trial.” Management Science 66 (2): 564–586. https://doi.org/10.1287/mnsc.2018.3249
• Camuffo, Arnaldo, Alfonso Gambardella, Danilo Messinese, Elena Novelli, Emilio Paolucci, and Chiara Spina. 2024. “A Scientific Approach to Entrepreneurial Decision-Making: Large-Scale Replication and Extension.” Strategic Management Journal 45 (6): 1209–1237. https://doi.org/10.1002/smj.3580
• Felin, Teppo, Alfonso Gambardella, Scott Stern, and Todd Zenger. 2020. “Lean Startup and the Business Model: Experimentation Revisited.” Long Range Planning 53 (4): 101889. https://doi.org/10.1016/j.lrp.2019.06.002
• Koning, Rembrand, Sharique Hasan, and Aaron Chatterji. 2022. “Experimentation and Start-up Performance: Evidence from A/B Testing.” Management Science 68 (9): 6434–6453. https://doi.org/10.1287/mnsc.2021.4209
• Leatherbee, Michael, and Riitta Katila. 2020. “The Lean Startup Method: Early-Stage Teams and Hypothesis-Based Probing of Business Ideas.” Strategic Entrepreneurship Journal 14 (4): 570–593. https://doi.org/10.1002/sej.1373
• Maier, Benjamin F., Ulf Aslak, Luca Fiaschi, Nina Rismal, Kemble Fletcher, Christian C. Luhmann, Robbie Dow, Kli Pappas, and Thomas V. Wiecki. 2025. “LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings.” arXiv:2510.08338. https://arxiv.org/abs/2510.08338
• Oetzel, Sebastian, and Tobias Maiberger. 2026. “How Price Sensitive Is AI? Benchmarking LLM-Generated Willingness to Pay in Pricing Analytics.” Journal of Marketing Analytics. https://doi.org/10.1057/s41270-026-00542-7
• Sarstedt, Marko, Susanne J. Adler, Lea Rau, and Bernd Schmitt. 2024. “Using Large Language Models to Generate Silicon Samples in Consumer and Marketing Research: Challenges, Opportunities, and Guidelines.” Psychology & Marketing 41 (6): 1254–1270. https://doi.org/10.1002/mar.21982
• Schmidt, Jonas, and Tammo H. A. Bijmolt. 2020. “Accurately Measuring Willingness to Pay for Consumer Goods: A Meta-Analysis of the Hypothetical Bias.” Journal of the Academy of Marketing Science 48 (3): 499–518. https://doi.org/10.1007/s11747-019-00666-6
• Wang, Angelina, Jamie Morgenstern, and John P. Dickerson. 2025. “Large Language Models That Replace Human Participants Can Harmfully Misportray and Flatten Identity Groups.” Nature Machine Intelligence 7 (3): 400–411. https://doi.org/10.1038/s42256-025-00986-z